Re: [PATCH v3 bpf-next 2/2] selftest/bpf: Implement sample UNIX domain socket iterator program.

[Yonghong Song <yhs@fb.com>]

On 8/4/21 12:08 AM, Kuniyuki Iwashima wrote:
> If there are no abstract sockets, this prog can output the same result
> compared to /proc/net/unix.
> 
>    # cat /sys/fs/bpf/unix | head -n 2
>    Num       RefCount Protocol Flags    Type St Inode Path
>    ffff9ab7122db000: 00000002 00000000 00010000 0001 01 10623 private/defer
> 
>    # cat /proc/net/unix | head -n 2
>    Num       RefCount Protocol Flags    Type St Inode Path
>    ffff9ab7122db000: 00000002 00000000 00010000 0001 01 10623 private/defer
> 
> According to the analysis by Yonghong Song (See the link), the BPF verifier
> cannot load the code in the comment to print the name of the abstract UNIX
> domain socket due to LLVM optimisation.  It can be uncommented once the
> LLVM code gen is improved.

I have pushed the llvm fix to llvm14 trunk 
(https://reviews.llvm.org/D107483), and filed a request to backport to 
llvm13 (https://bugs.llvm.org/show_bug.cgi?id=51363), could you in the 
next revision uncomment the "for" loop code and tested it with latest 
llvm trunk compiler? Please also add an entry in selftests/bpf/README.rst
to mention the llvm commit https://reviews.llvm.org/D107483 is needed
for bpf_iter unix_socket selftest, otherwise, they will see an error
like ...

> 
> Link: https://lore.kernel.org/netdev/1994df05-8f01-371f-3c3b-d33d7836878c@fb.com/
> Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
> ---
>   .../selftests/bpf/prog_tests/bpf_iter.c       | 16 ++++
>   tools/testing/selftests/bpf/progs/bpf_iter.h  |  8 ++
>   .../selftests/bpf/progs/bpf_iter_unix.c       | 86 +++++++++++++++++++
>   .../selftests/bpf/progs/bpf_tracing_net.h     |  4 +
>   4 files changed, 114 insertions(+)
>   create mode 100644 tools/testing/selftests/bpf/progs/bpf_iter_unix.c
> 
> diff --git a/tools/testing/selftests/bpf/prog_tests/bpf_iter.c b/tools/testing/selftests/bpf/prog_tests/bpf_iter.c
> index 1f1aade56504..77ac24b191d4 100644
> --- a/tools/testing/selftests/bpf/prog_tests/bpf_iter.c
> +++ b/tools/testing/selftests/bpf/prog_tests/bpf_iter.c
> @@ -13,6 +13,7 @@
>   #include "bpf_iter_tcp6.skel.h"
>   #include "bpf_iter_udp4.skel.h"
>   #include "bpf_iter_udp6.skel.h"
> +#include "bpf_iter_unix.skel.h"
>   #include "bpf_iter_test_kern1.skel.h"
>   #include "bpf_iter_test_kern2.skel.h"
>   #include "bpf_iter_test_kern3.skel.h"
> @@ -313,6 +314,19 @@ static void test_udp6(void)
>   	bpf_iter_udp6__destroy(skel);
>   }
>   
> +static void test_unix(void)
> +{
> +	struct bpf_iter_unix *skel;
> +
> +	skel = bpf_iter_unix__open_and_load();
> +	if (!ASSERT_OK_PTR(skel, "bpf_iter_unix__open_and_load"))
> +		return;
> +
> +	do_dummy_read(skel->progs.dump_unix);
> +
> +	bpf_iter_unix__destroy(skel);
> +}
> +
[...]
> +	if (unix_sk->addr) {
> +		if (!UNIX_ABSTRACT(unix_sk)) {
> +			BPF_SEQ_PRINTF(seq, " %s", unix_sk->addr->name->sun_path);
> +		} else {
> +			BPF_SEQ_PRINTF(seq, " @");
> +
> +			/* The name of the abstract UNIX domain socket starts
> +			 * with '\0' and can contain '\0'.  The null bytes
> +			 * should be escaped as done in unix_seq_show().
> +			 * However, the BPF verifier cannot load the code below
> +			 * because of the optimisation by LLVM.  So, print only
> +			 * the first escaped byte here for now.  Once LLVM code
> +			 * gen is improved, remove the BPF_SEQ_PRINTF() above
> +			 * and uncomment the code below.
> +			 *
> +			 * int i, len;
> +			 *
> +			 * len = unix_sk->addr->len - sizeof(short);
> +			 *
> +			 * BPF_SEQ_PRINTF(seq, " @");
> +			 *
> +			 * // unix_mkname() tests this upper bound.
> +			 * if (len < sizeof(struct sockaddr_un))
> +			 *	for (i = 1 ; i < len; i++)
> +			 *		BPF_SEQ_PRINTF(seq, "%c",
> +			 *			       unix_sk->addr->name->sun_path[i] ?:
> +			 *			       '@');
> +			 */
> +		}
> +	}
> +
[...]